Utility of Different Exposure Metrics used in Epidemiological - PowerPoint PPT Presentation

Utility of Different Exposure Metrics used in Epidemiological Studies of Air Pollution Halûk Özkaynak US EPA, Office of Research and Development National Exposure Research Laboratory, RTP, NC Presented at the NYSERDA EMEP 2011 Conference Albany, NY November 15, 2011

Estimated Effects of Ambient PM 2.5 on Acute Mortality in the US * Community-specific estimates of the percent increase in respiratory mortality with a 10 µ g/m 3 increase in the previous day's PM 2.5 concentrations 25 20 15 10 Percent Increase 5 0 -5 -10 -15 -20 Las Vegas Los Angeles Phoenix Birmingham Boston Chicago Cincinnati Cleveland Columbus Dallas Detroit Fresno Houston Indianapolis Manhattan Memphis Milwaukee Minneapolis Palm Beach Philadelphia Pittsburgh Riverside Sacramento San Diego Seattle Tampa Washington DC Community ▪ represents estimates; lines around ▪ are 95% confidence interval *Source of data : Franklin et al. 2007)

Need for Better Exposure Characterization in Air Pollution Health Studies • Numerous epidemiologic studies have used measurements from central-site ambient monitors as surrogates of personal exposures to air pollution • Central-site monitors may not account for:  spatial and temporal heterogeneity of urban air ambient pollution  human activity patterns  infiltration of ambient pollutants indoors  contributions of indoor sources that may be effect modifiers • Central-site are especially problematic for certain PM components and species (e.g., EC, OC, coarse, ultrafine) that exhibit significant spatial heterogeneity • A number of enhanced exposure assessment approaches have recently been developed and applied in the investigation of air pollution health effects by EPA and collaborating academic institutions

Exposure Metrics Considered by Health Studies Tiers of Exposure Metrics Input data Complexity Reliability vs. Ambient Monitoring Data: Monitoring Data Central Site or Interpolated Uncertainty Land-Use Regression Monitoring Data Emissions Data or Intake Fraction Models Land-Use/Topography Air Quality Modeling Emissions Data Meteorological Data (CMAQ, AERMOD, hybrid) Land-Use/Topography Monitoring Data Statistical/Hybrid Modeling Emissions Data Meteorological Data (Data blending) Land-Use/Satellite Monitoring Data Emissions Data Exposure Modeling Meteorological Data Land-Use/Topography (SHEDS, APEX) Personal Behavior/Time Activity Microenvironmental Characteristics Health data analysis Epidemiological statistical models: log(E( Y kt )) = α + β exposure metric kt + ∑ k γ k area kt + … other covariates

Stochastic Human Exposure and Dose Simulation (SHEDS) Model for Air Pollutants

Comparison of Effects per IQR Unit Change in Ozone Concentration vs. Exposure on Respiratory Hospital Admissions in NYC (Jones et al. 2011) 1.10 Ozone concentration Exposure (Daily 8hr max) (mean of daily 8 hr max) HR for Admission (95% CI) IQR: 25.3 IQR: 6.03 1.041 1.022 1.033 1.020 1.029 1.017 1.05 1.024 1.005 1.002 1.008 1.00 0 1 2 3 4 0 1 2 3 4 Lag days 0.95 Conditional logistic regression model adjusted for categorical mean UAT.

Tier 3 - SHEDS/APP: Relative odds of a transmural infarction in NJ associated with each IQR increase in 1.60 PM 2.5 concentration in the 24 hours before ER arrival* 1.50 (n=1550) Tier 1 - PM2.5 - TEOM Tier 3 - PM2.5 - Tier3 - 1 value per zipcode 1.40 AIC 1.30 4374.8 4373.9 AIC = 4375.5 1.20 Odds Ratio 1.10 1.00 0.90 PM 2.5 Tier3 Each PM 2.5 metric (1 value per zipcode) modeled separately 0.80 & PM 2.5 TEOM in model together, using 0.70 Z score methods 0.60 *Source : Turpin et al. 2011

Modeled 4-year average NO x and PM 2.5 concentrations in Atlanta: a) regional background and b) hybrid (regional combined with local) !999-2002

Results of the epidemiologic analysis of emergency department data in Atlanta for a) respiratory diseases and b) cardiovascular diseases (Özkaynak et al. 2011)

Associations between 24h NOx/CO and Asthma ED Visits In Atlanta (Sarnat et al. 2011) 1.04 Relative Risk (95% CI) per IQR increase in Pollutant Metric 1.03 1.02 1.01 1.00 0.99 0.98 0.97 0.96 CS BG AM HyB APEX APEX CS BG AM HyB APEX APEX CS24 BG24 AERMOD HYBRID APEXP50 APEXP95 CS24 BG24 AERMOD HYBRID APEXP50 APEXP95 Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze Asthma/Wheeze p50 p95 p50 p95 CO NOx

Overall Summary of Findings • Observed RRs differ by metric, pollutant and study design - Measurement error is present in every metric - Effects of error on risk estimates vary by type of exposure error (Goldman et al 2011) - For time-series studies ambient concentrations may serve as an appropriate exposure surrogate -For cohort studies or mixed spatio-temporal study designs (as shown in the Atlanta analysis) the use of more refined exposure surrogates than the conventionally used ambient monitoring data may boost study power, reduce exposure prediction errors and strengthen the estimated associations between air pollution and health data • CO, NOx  Asthma ED Associations Varied by Metric Choice • Model-based estimate higher and significant compared to central site estimate • Consistent with a priori expectations for spatially-heterogeneous pollutants - Suggests that accounting for spatiotemporal distribution of pollutants may be important for timeseries studies - May indicate reduced measurement error for these pollutants • Ozone, PM 2.5  Magnitudes of Association with Daily Mortality, MI, Respiratory Hospitalization s and Emergency Department Visits Fairly Robust to Metric • Interpretation of findings similar regardless of exposure assignment approach • For homogenous pollutants, spatiotemporal models may add little to explaining variability • Slightly lower RRs for the modeled O 3 personal estimates compared to ambient • Possible that potential for exposure model misspecification may re-introduce error in the epidemiological analysis results using modeled exposures

Research Needs Type of an epidemiologic study design influences spatio-temporal resolution needs of exposure data or its surrogates used for health effects research (one size does not fit all in terms of optimally assigning exposures). Key information gaps: 1) When do more refined estimates of exposure provide more information than the central-site monitor, by: 1) Type of study (e.g. case-crossover vs. cohort); 2) Acute vs. chronic exposures/effects, 3) Spatial vs. temporal variability of pollutant of interest It is important to better understand the sources and factors influencing uncertainties in ambient pollution epidemiology analyses as well as compounding of errors as exposure metrics are refined. Key information gaps: 2) What is the best way to apply distributional exposure estimates? 3) How much infiltration, activity patterns, local source emissions and pollution composition account for the predicted variability in the exposure and effect estimates? 4) How to incorporate multipollutant considerations in modeling exposures and epidemiological analyses, since appropriate selection of exposure surrogates become more complicated due to pollutant-specific relationships with their exposure surrogates and the underlying covariance structure among the ambient pollutant concentrations (i.e.., exposure misclassification concerns vs. statistical collinearity issues)?

Acknowledgements • Vlad Isakov(EPA) • Lisa Baxter (EPA) • Val Garcia (EPA) • Janet Burke (EPA) • ST Rao (EPA) • Barbara Turpin (Rutgers) • David Rich (University of Rochester) • Jeremy and Stefanie Sarnat (Emory) • Jim Mulholland (Georgia Tech) • Rena Jones, Shao Lin (NYSDOH) • Chris Frey, Montse Fuentes (NCSU)